Incremental Evolving Domain Adaptation

Almost all of the existing domain adaptation methods assume that all test data belong to a single stationary target distribution. However, in many real world applications, data arrive sequentially and the data distribution is continuously evolving. In this paper, we tackle the problem of adaptation to a continuously evolving target domain that has been recently introduced. We assume that the available data for the source domain are labeled but the examples of the target domain can be unlabeled and arrive sequentially. Moreover, the distribution of the target domain can evolve continuously over time. We propose the Evolving Domain Adaptation (EDA) method that first finds a new feature space in which the source domain and the current target domain are approximately indistinguishable. Therefore, source and target domain data are similarly distributed in the new feature space and we use a semi-supervised classification method to utilize both the unlabeled data of the target domain and the labeled data of the source domain. Since test data arrives sequentially, we propose an incremental approach both for finding the new feature space and for semi-supervised classification. Experiments on several real datasets demonstrate the superiority of our proposed method in comparison to the other recent methods.